Graph Neural Network Modelling of Pre-Diagnostic Serum Proteomics for Pancreatic Cancer Detection

Abstract

Background/Objectives: Pancreatic ductal adenocarcinoma (PDAC) has exceptionally high mortality, largely because of late diagnosis, while established blood biomarkers such as CA19-9 have limited sensitivity for early detection. We evaluated whether modelling serum proteomic measurements as supervised pairwise statistical graphs using synolitic graph neural networks (SGNNs) could provide a useful framework for pre-diagnostic PDAC classification. Methods: We analysed 217 pre-diagnostic serum samples from UKCTOCS, comprising 100 PDAC samples from 75 women and 117 control samples, profiled for 97 proteins. Samples were divided into a development set collected 0–1 year before diagnosis (n = 110) and a temporal holdout set collected 1–2 years before diagnosis (n = 107); 25 PDAC participants contributed samples to both windows. The final SGNN used fold-internal mutual-information selection of 30 proteins, pairwise RBF-SVM-derived edge information, minimum-connected sparsification, and a 10-member GATv2 ensemble per cross-validation fold. Conventional machine-learning comparators were tuned using development data only. Results: The SGNN achieved a mean 5-fold development ROC-AUC of 75.33 ± 10.43%, lower than the tuned Random Forest (81.67%); the other tuned comparators achieved 77.67% (XGBoost), 77.17% (logistic regression), and 73.17% (SVM). Averaging predictions across the five-fold-specific SGNN ensembles yielded a temporal-holdout ROC-AUC of 64.35% (95% CI 52.6–74.6%). In the participant-independent subset of 25 PDAC cases not represented in development and 57 controls, ROC-AUC was 66.04% (95% CI 52.56–78.88%). Conclusions: SGNNs provide a feasible graph-based representation of pre-diagnostic proteomic data but did not outperform optimised conventional machine-learning methods within development cross-validation. The temporal and participant-independent performance estimates warrant further investigation in larger fully independent cohorts. A graph-specific predictive advantage was not directly tested—no such advantage was demonstrated in this dataset—and the clinical utility has not been established.